Founder & CEO of Discera.
Routing writing
I build an AI router that sends each message to the model best suited for it. Scroll, and watch it work:
>
scroll to route ↓
The motto
Every task has a different winner.
What my research company, Noesira, is built on.
Inside an AI · illustrated tour
01 / 07
Noesira field trip · illustrated tour
Inside
an AI
An illustrated journey from computing hardware to a model’s reply, in six stops. Keep scrolling.
Ask one of the big AI services a question, and the work starts in a building like this: a data center packed with computers running day and night.
Fig. 03·1 — data center
Inside: rack after rack of GPUs — chips built to do the same kind of math over and over, at enormous scale.
Fig. 03·2 — GPU rack
One GPU spreads the work across thousands of small cores, all computing at the same time.
Fig. 03·3 — one GPU
Together they run a model: a vast web of connections, each one adjusted during training on enormous amounts of data.
Fig. 03·4 — the model
Each connection carries a number the model learned — a weight. Your prompt is turned into tokens and processed through billions of them.
Fig. 03·5 — weights
Out comes the reply, one token at a time. Noesira’s question: which model does it best?
Fig. 03·6 — the answer
I built Discera because no single model is best at everything. It's one prompt box that reads your task and sends it to the model that handles it best.
One prompt in, the right model out
Fig. 02
Discera reports 94.1% average routing accuracy across three 147-question test sets, versus 73.7% for the original router it was measured against on the same benchmark. As published on discera.chat
Visit Disceradiscera.chat (access is limited by code)↗I started Noesira to rank AI models by what they're actually best at: writing, coding, reasoning, research, routing. It's compiled from public benchmarks and arenas, and published with the sources attached.